What Is an MCP Server?
A plain-English breakdown of the Model Context Protocol's architecture, how it differs from a regular API, and what it looks like running in production at 242 tools.
By Austin Archuleta, Founder & AI Solutions Architect — Published August 10, 2026
Quick Answer
What Is an MCP Server?
An MCP server is a program that exposes a defined set of tools, data, and prompts to AI applications over the Model Context Protocol—an open standard Anthropic released in November 2024. It lets an AI agent discover what actions are available (update a CRM record, query a database, send an email) and call them directly, instead of a developer writing custom, model-specific integration code for every connection.
Last updated: August 10, 2026
The Three Pieces: Host, Client, Server
MCP defines three roles, and understanding them is the fastest way to understand what an MCP server actually is—and isn't. When someone says “MCP server,” they mean specifically the third piece:
Host
The AI application the person actually uses—Claude Desktop, ChatGPT, or a custom agent. The host manages the overall conversation and decides when a tool is needed.
Client
A connector that lives inside the host and manages one stateful, 1:1 connection to a single server—handling the protocol handshake, tool discovery, and message routing for that connection.
Server
A lightweight program that exposes a defined set of tools, resources, and prompts. It executes the real action—reading a database, calling an API—and returns a structured result to the client.
In practice: a user asks Claude Desktop (the host) to check on a lead's status. The host's client sends a request over its connection to the CRM's MCP server. The server looks up the contact in the real CRM, formats the result, and sends it back. The model incorporates that structured result into its answer—it never touches the CRM's database or credentials directly, and the server never has to know anything about how Claude generates text.
One host can run multiple clients at once, each maintaining a separate connection to a different server—one for the CRM, one for email, one for a database. That's what lets a single AI agent operate across several systems in one conversation.
MCP Server vs. API vs. Plugin vs. RPA
“Isn't this just an API?” is the most common question. Here's where an MCP server actually sits relative to the tools it gets compared to.
| Factor | MCP Server | Traditional API | Chat Plugin/GPT | RPA Bot |
|---|---|---|---|---|
| Built for | AI models to discover & call at runtime | Developers to read docs & code against | Browser extensions and third-party GPT integrations | Recorded UI clicks, no model involved |
| Discovery | Model reads tool descriptions automatically | Manual—read the docs, write the client | Manual, per-platform submission | None—hardcoded steps |
| Works across AI models | Yes—any MCP-compatible client | N/A, not model-specific | No—locked to one platform's plugin system | N/A, not AI-driven |
| Handles ambiguous requests | Yes—model decides which tool fits | No—fixed integration logic | Limited—platform-defined triggers | No—breaks on any UI change |
| Typical use | AI agents taking judgment-based action | System-to-system, deterministic sync | In-chat plugins for one AI product | Repetitive, unchanging desktop tasks |
An MCP server usually wraps a traditional API rather than replacing it—the API does the work, the MCP layer makes it something an AI model can find and call correctly. For a deeper look at when to reach for each option, see our MCP server development service page.
A Production Example: 242 Tools, Two Transports
Definitions only go so far—here's what an MCP server looks like once it's actually running in production, not a demo. We built Follow Up Ace, an AI layer for real estate teams on Follow Up Boss CRM, around exactly the architecture above—but real systems rarely fit one transport.
Follow Up Ace runs 242 tool definitions on a stdio MCP server—the local, subprocess-based transport used by Claude Desktop—exposing real CRM operations: creating contacts, logging activity, drafting follow-ups, booking tasks. Alongside it, a separate 123-tool HTTP surface serves hosted connectors that need a remote, multi-user connection instead of a local subprocess, and a desktop extension proxies through server-side OAuth so an individual user's credentials never have to be embedded in a client config file.
That split is the practical answer to a question the architecture section glosses over: real MCP deployments often run more than one server, or more than one transport for the same tool surface, because different clients connect differently. The protocol handles the discovery and calling convention; the engineering work is in getting the transport, auth, and tool descriptions right for each one. See the full Follow Up Ace case study for how the rest of the system—five retrained scoring models, a BigQuery warehouse, an ElevenLabs voice agent—sits behind that MCP layer.
Common Questions
What is the difference between an API and an MCP server?
A regular API is built for a developer who already knows the exact endpoint, parameters, and response shape—the documentation lives in a separate doc a human reads once and codes against. An MCP server is built for an AI model to discover at runtime: each tool ships with a natural-language description the model reads to decide whether and how to call it, plus a typed parameter schema it fills in automatically. An MCP server can wrap a regular API—it's a discovery and calling convention layered on top, not a replacement for the underlying system.
Why would I need an MCP server?
You need one when you want an AI agent to take real action in your existing systems—updating a CRM record, querying a database, sending an email—rather than just generating text about them. Without one, connecting an AI model to your tools means custom, model-specific integration code that has to be rebuilt every time you add a tool or switch models. An MCP server decouples the two: build the tool surface once, and Claude, ChatGPT, or any other compatible agent can call it.
Is an MCP server a real server?
It depends on the transport. A local (stdio) MCP server runs as a subprocess on the same machine as the AI client—no network, no separate hosting, the standard setup for Claude Desktop. A remote MCP server is a real, standalone server in the traditional sense: a hosted web service over HTTP that multiple users or agents can connect to concurrently, deployed the same way you'd deploy any API—a container, a serverless function, or a Node.js/Python process behind a reverse proxy.
Does ChatGPT use MCP?
Yes. OpenAI adopted MCP across its products, including the ChatGPT desktop app, in March 2025—about four months after Anthropic introduced the protocol. That made MCP a cross-vendor standard rather than a Claude-only feature, which is the main reason it's worth building an MCP server instead of a model-specific integration: the same server works with Claude, ChatGPT, and any other agent that implements the spec.
Is MCP the same as function calling?
No, but they work together. Function calling (also called tool calling) is a capability built into the AI model itself—its ability to look at a list of available tools and emit a structured request to call one. MCP is the standardized protocol that gets that tool list, and the call, between the model and whatever system actually performs the action. Function calling is what the model does; MCP is how the tools get to the model in the first place, in a form any compatible model can use without custom wiring.
Who created MCP and who maintains it now?
Anthropic created and open-sourced the Model Context Protocol in November 2024 as a specification, not a proprietary product. In December 2025, Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation—moving it to vendor-neutral governance so its roadmap doesn't depend on any single AI company.
Need an MCP Server Built for Production?
We design and build production MCP servers—authentication, rate limiting, multi-tenant isolation, and tool descriptions precise enough for a model to call reliably. Book a free scoping call to map your tool surface.